A Multi-strategy Region Proposal Network

作者：

Highlights：

• MSRPN is proposed based on the RPN of Faster R-CNN.

• A novel skip-layer connection network and improved anchor boxes are designed.

• A novel unified classification and regression layer is introduced.

• The bounding box regression part of multi-task loss function is improved.

• MSRPN achieves excellent experiment results on object detection.

摘要

•MSRPN is proposed based on the RPN of Faster R-CNN.•A novel skip-layer connection network and improved anchor boxes are designed.•A novel unified classification and regression layer is introduced.•The bounding box regression part of multi-task loss function is improved.•MSRPN achieves excellent experiment results on object detection.

论文关键词：Object detection,Region proposal generation,Convolutional neural network,Classification

论文评审过程：Received 4 January 2018, Revised 18 June 2018, Accepted 19 June 2018, Available online 28 June 2018, Version of Record 5 July 2018.

论文官网地址：https://doi.org/10.1016/j.eswa.2018.06.043